Today's Fine-tuning & Training: Fastest-Growing Projects — April 18, 2026
This week, the Fine-tuning & Training space on GitHub is dominated by innovations in large language model (LLM) optimization and compression. Repositories focused on efficient training methods, such as sparse retrieval PEFT and near-optimal KV cache quantization, are gaining traction among developers. Meanwhile, implementations of cutting-edge research papers, like Google's TurboQuant, continue to attract attention.
Facebookresearch/tribev2, with a Growth Score of 63.28 and 1,883 stars, is a prominent repository containing the code to train and evaluate TRIBE v2, a multimodal model for brain response prediction. Its growth can be attributed to the increasing interest in multimodal AI research and its potential applications.
QingGo/engram-peft boasts an impressive Growth Score of 34.25 and 24 stars, despite being a relatively new repository. This unofficial implementation of DeepSeek Engram injects high-capacity conditional memory into LLMs via sparse retrieval PEFT, making it an attractive solution for developers seeking to enhance their models' capabilities.
0xSero/turboquant has garnered significant attention with its near-optimal KV cache quantization for LLM inference, earning a Growth Score of 30.60 and 1,079 stars. Its growth is driven by the demand for efficient training methods that can reduce computational costs without sacrificing model performance.
Tonbistudio/turboquant-pytorch offers a PyTorch implementation of Google's TurboQuant, achieving 5x compression at 3-bit with 99.5% attention fidelity. With a Growth Score of 28.25 and 942 stars, this repository is popular among developers seeking to integrate TurboQuant into their workflows.
WillowHe/EvoOpt_oppangu_optimization_model provides solutions for fine-tuning large language models in operations research optimization tasks using Openpangu-7B as the base model. Although its Growth Score of 9.37 and 335 stars are relatively modest, this repository addresses a specific niche that is gaining interest.
OnlyTerp/turboquant presents another open-source implementation of Google TurboQuant, with a focus on near-optimal KV cache compression for LLM inference. With a Growth Score of 6.92 and 53 stars, its growth can be attributed to the ongoing demand for efficient training methods.
Dynamis-Labs/spectralquant proposes an alternative approach to breaking TurboQuant's compression limit via spectral structure, achieving impressive results. Its Growth Score of 6.54 and 110 stars indicate a growing interest in exploring new optimization techniques.
Mintzs/oogaboogalm explores the idea of fine-tuning AI models to reduce token use by incorporating caveman system prompts and skills into the model itself. Although its Growth Score of 7.19 and 40 stars are relatively low, this repository represents an innovative approach to optimizing LLMs.
Other notable repositories in the Fine-tuning & Training space include SUM-INNOVATION/RUMUS, a Rust-based framework for training neural networks (Growth Score: 8.47, Stars: 65), and 917017420/codex-register-fix, which offers an openAI registration learning project based on cnlimiter/codex-manager (Growth Score: 7.00, Stars: 70).
Facebookresearch/tribev2, with a Growth Score of 63.28 and 1,883 stars, is a prominent repository containing the code to train and evaluate TRIBE v2, a multimodal model for brain response prediction. Its growth can be attributed to the increasing interest in multimodal AI research and its potential applications.
QingGo/engram-peft boasts an impressive Growth Score of 34.25 and 24 stars, despite being a relatively new repository. This unofficial implementation of DeepSeek Engram injects high-capacity conditional memory into LLMs via sparse retrieval PEFT, making it an attractive solution for developers seeking to enhance their models' capabilities.
0xSero/turboquant has garnered significant attention with its near-optimal KV cache quantization for LLM inference, earning a Growth Score of 30.60 and 1,079 stars. Its growth is driven by the demand for efficient training methods that can reduce computational costs without sacrificing model performance.
Tonbistudio/turboquant-pytorch offers a PyTorch implementation of Google's TurboQuant, achieving 5x compression at 3-bit with 99.5% attention fidelity. With a Growth Score of 28.25 and 942 stars, this repository is popular among developers seeking to integrate TurboQuant into their workflows.
WillowHe/EvoOpt_oppangu_optimization_model provides solutions for fine-tuning large language models in operations research optimization tasks using Openpangu-7B as the base model. Although its Growth Score of 9.37 and 335 stars are relatively modest, this repository addresses a specific niche that is gaining interest.
OnlyTerp/turboquant presents another open-source implementation of Google TurboQuant, with a focus on near-optimal KV cache compression for LLM inference. With a Growth Score of 6.92 and 53 stars, its growth can be attributed to the ongoing demand for efficient training methods.
Dynamis-Labs/spectralquant proposes an alternative approach to breaking TurboQuant's compression limit via spectral structure, achieving impressive results. Its Growth Score of 6.54 and 110 stars indicate a growing interest in exploring new optimization techniques.
Mintzs/oogaboogalm explores the idea of fine-tuning AI models to reduce token use by incorporating caveman system prompts and skills into the model itself. Although its Growth Score of 7.19 and 40 stars are relatively low, this repository represents an innovative approach to optimizing LLMs.
Other notable repositories in the Fine-tuning & Training space include SUM-INNOVATION/RUMUS, a Rust-based framework for training neural networks (Growth Score: 8.47, Stars: 65), and 917017420/codex-register-fix, which offers an openAI registration learning project based on cnlimiter/codex-manager (Growth Score: 7.00, Stars: 70).